Guardrails before autonomy
A model that can spend money needs hard limits in code, not polite instructions in a prompt. Caps, windows and forbidden actions come first; autonomy is a switch you turn on.
About
A control layer for paid acquisition
AdBeeCore is built by a team that ran paid acquisition in-house for apps, websites and service businesses: dozens of accounts across Meta, Google Ads and Apple Search Ads, four consoles, and a spreadsheet that was always one day behind.
Facts over forecasts
We show what the networks reported and what changed in the account. When data is missing, the product says so instead of interpolating.
Every number traceable
Each metric can be traced to a source and each change to an author. If a number looks wrong, that is a bug, not an opinion.
Boring reliability
Attribution windows, rate limits and retries are not exotic features — they are the whole job. One failed account must never break a collection cycle.
How it is built
The same components you would pick for a serious internal tool: typed end to end, boringly observable.
- Runtime
- Bun and ElysiaJS for the API, Astro with React islands for the console — type-safe from the database to the button.
- Data
- PostgreSQL for the source of truth, Redis for locks and cache, Qdrant for agent memory and retrieval.
- Adapters
- One AdProvider interface over Google, Meta, Apple and TikTok, with per-network rate limiting, retries and typed errors.
- Agents
- Claude for reasoning, with a strict tool contract: every write goes through Guard, a change log entry and — where needed — a human confirmation.
See it on your data
Free sign-up, one account, observe mode. The fastest way to judge a control layer is to watch it find something you missed.